Category Intelligent Software>Neural Network Systems/Tools

Abstract ThinksPro is a complete neural network application development environment that can be used to develop and deploy applications ranging from controllers to mainframes. Product provides visualization tools, algorithms and Dynamic Link Libraries (DLLs) so that trained neural networks can be embedded into your custom applications. It contains some of the most advanced algorithms and capabilities available.

ThinksPro provides multiple views into how neural networks learn and work. ThinksPro takes you into the world of reinforcement learning via its EvalNet User Defined Grading Function and four (4) graded learning algorithms, including the Temporal Difference Algorithm. Product can be used for forecasting, prediction, financial analysis, classification, function approximation, decision making, sensor interpretation, and control applications.

Features/Capablities include:

Windows Dynamic Link Library (DLL) - Allows the professional user to build custom Windows, Visual Basic and Excel applications.

Sensitivity Analysis of Input Variables - Sensitivity Analysis facilitates the determination of the relative importance of the neural network inputs. Input disabling allows you to turn off unimportant inputs without reloading or rebuilding the training data.

EvalNet User Defined Grading (DLL) - This allows the user to interactively test the network during the learning process. This also allows ThinksPro to train networks for complex control applications or to maximize return on investment or any cost function in financial applications. This feature brings the full power of the ThinksPro visualization tools to your custom application.

ThinksPro provides maximum flexibility in seven (7) categories:

a) Preprocessing Functions; b) Interconnection Architectures; c) Processing Element Summation Functions; d) Processing Element Transfer Functions; e) Learning Algorithms; f) Learning Error Criteria; g) Visualization Tools.

a) Preprocessing Functions - Product provides special input layer preprocessing to simplify working with all forms of data - 1) Min/Max and Mean/StdDev Input Preprocessing; 2) Sum Inputs to 1 Normalization; 3) Sum of Squares to 1 Normalization; 4) Time Series Window Preprocessor; 5) Input Disabling and Input Sensitivity Analysis.

b) Interconnection Architectures - Processing elements can be connected in networks up to five (5) layers - 1) Multilayer Normal Feed Forward; 2) Multilayer Full Feed Forward; 3) Cascade; 4) Total Recurrent; 5) Jordan Recurrent; 6) Elman Recurrent; Prior Recurrent; 7) Cascade Recurrent.

c) Processing Element Summation Functions - The function can be chosen independently on each layer of the network - 1) Dot Product; 2) Quadratic Sum; 3) L1 Distance; 4) L2 Distance; 5) Radial Basis Function; 6) Sigma-Pi; 7) General Regression Neural Network (GRNN) Sum.

d) Processing Element Transfer Functions - ThinksPro offers greater efficiency and training with 14 transfer functions - 1) Sigmoid; 2) Bipolar Sigmoid; 3) Arctan; 4) Bipolar Arctan; 5) Sin; 6) Bipolar Sin; 7) Threshold Linear; 8) Bipolar Threshold Linear; 9) Threshold; 10) Bipolar Threshold; 11) Linear; 12) Gaussian; 13) Cauchy; 14) Winner Take All.

e) Learning Algorithms - Product supports fifteen (15) industry standard algorithms - 1) Back Propagation; 2) Quick Propagation; 3) Jacob's Enhanced Back Propagation; 4) Cascade Correlation Learning; 5) Conjugate Gradients; 6) Levenberg-Marquardt; 7) Recurrent Back Propagation; 8) Kohonen Winner Take All; 9) Kohonen Learning Vector Quantization; 10) Solis & Wets Simulated Annealing (graded); 11) Simplex Simulated Annealing (graded); 12) Powell’s Method Without Derivatives (graded); 13) Probabilistic Neural Network; 14) General Regression Neural Network; 15) Temporal Difference Learning.

f) Learning Error Criteria - Control the network’s sensitivity to outliers in the training data with five (5) error function options - 1) Mean Squared Error; 2) Mean Absolute Error; 3) Mean Fourth Power Error; 4) Hyperbolic Square Error; 5) Bipolar Hyperbolic Square Error; 6) Classification Error; 7) User Defined Grading Function.

g) Visualization Tools - Ability to examine the neural network inputs, weights and states during training with time series, Hinton and color diagrams.

System Requirements

Windows 2000 or better.

Also has executables and shared objects for Linux on an Intel processor.

For large training runs we recommend at least 1G of ram.


Manufacturer Web Site Logical Designs Consulting, Inc.

Price $1,495.00

G6G Abstract Number 20052

G6G Manufacturer Number 101705